Accurate small-sample prediction of ultra-fast laser cutting quality for quartz pendulous via WGAN-augmented ANN

The quartz pendulous, serving as the core component of the quartz flexible accelerometer, requires controlled fabrication to enhance accelerometer performance and enable effective error compensation. However, achieving rapid and controlled machining quality presents considerable challenges due to the complex interactions between laser parameters during ultra-fast laser processing. While artificial neural network (ANN) offers potential for nonlinear optimization, they typically demand large datasets to ensure prediction accuracy. To overcome the limitation imposed by the need for extensive sample data, an orthogonal experimental design was adopted to obtain 16 original datasets. An approach using Wasserstein generative adversarial networks (WGANs)-augmented ANN was applied, where high-fidelity synthetic data generated by the WGAN augmented the dataset. Training the ANN on this augmented data enabled the precise establishment of relationships between process parameters and key cutting quality metrics-taper angle and surface roughness. On the training set and the test set, the taper angle <i>R</i>² increased by 10.25% and 5.85%, respectively. The roughness <i>R</i>² shows a significant gain, increasing by 209.32% on the training set and by 38.80% on the test set. Rigorous validation with what we believe are five novel process parameter sets demonstrated robust generalization, yielding average relative errors of 3.682% for taper angle and 1.655% for roughness against experimental measurements. These results confirm the method's efficacy for accurate quality prediction using minimal data, enabling accelerated process optimization and precision control for ultra-fast laser machining.